Block Unveils Dessa, an On-Device Full-Duplex Speech Model
Jack Dorsey's Block has unveiled Dessa, a native speech-to-speech AI model trained on a single GPU that runs locally and replicates nuanced human vocal expressions.
Jack Dorsey's financial technology firm Block has introduced Dessa, an experimental voice model developed by the company's research laboratory. Built to run locally on consumer-grade hardware, the system is designed to replicate human vocal interactions with a high degree of authenticity, incorporating conversational elements such as pauses, laughter, and crying.
Dessa functions as a full-duplex voice model, enabling it to process incoming audio and generate speech simultaneously. This operational design mirrors capabilities introduced earlier this year in OpenAI's GPT-Live system. However, while mainstream interactive audio setups typically route incoming speech through a transcription pipeline into a large language model before rendering text back into audio, Block claims Dessa's conversational knowledge and responses originate directly from its underlying conversational training data.
Moving Past Text-Based Audio Pipelines
According to Block, conventional voice systems frequently deliver rigid, unnatural interactions because they rely on intermediate text representations that discard non-verbal context. Existing market alternatives, including ElevenLabs v4 and Fish Audio Drama 3, often implement voice tags to instruct speech engines on how to apply emotional inflection. Block asserts that Dessa integrates emotional expression directly into the model architecture rather than relying on external steering tags.
In a blog post explaining the technical direction, the company argued that traditional voice systems diminish the user experience by stripping away critical acoustic context: "Interacting with current voice AI can just plain suck. It turns your voice into text and throws away an incredible amount of information text alone can’t carry. The most powerful AIs in the world can’t hear how you feel. We think that’s ridiculous, leads to a worse experience, and is potentially dangerous."
Block stated that the model was trained entirely on human speech recordings utilizing a single NVIDIA GPU. Demonstrations published by researchers highlighted the model adjusting its cadence, voice profile, and emotional state across diverse contexts. In one shared test, the agent sounded audibly fatigued when mentioning that it was tired from work. In another instance, the model sounded emotionally overwhelmed to the point of nearly breaking down while expressing gratitude for being made comfortable. A separate demonstration showed a Block team member playing an interactive game directly with the assistant.
Commenting on the release, Dorsey noted: "We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence." Ragavan Thurairatnam, who shared technical highlights of the system, described the release as a significant milestone, calling it "a GPT3 moment for speech" that achieves full-duplex speech-to-speech interaction on consumer hardware.
Hardware Demands and Research Access
Despite its expressive range, Dessa does not require specialized enterprise data center clusters for inference. Block confirmed that the system is engineered to operate locally on personal workstations, requiring a consumer-grade computer such as an Apple MacBook M4 Max or higher.
At present, Block has not released Dessa to the general public. Instead, the company has issued a call for safety researchers to evaluate the architecture. Because human-like vocal inflection, pauses, and realistic emotional expression introduce significant risks regarding social engineering, fraud, and user deception, the testing phase is expected to identify vulnerabilities and assess how the model behaves under adversarial scenarios.
What it means for developers
For engineering teams building conversational software, Dessa signals an architectural shift away from cascaded pipelines—which chain speech-to-text, large language models, and text-to-speech engines together—toward native speech-to-speech systems. By processing raw audio end-to-end, full-duplex models eliminate latency bottlenecks and retain conversational subtleties like interruptions and vocal hesitation.
The ability to execute full-duplex models on hardware like the MacBook M4 Max also highlights potential cost efficiencies for commercial applications. Block already deploys voice AI for ordering workflows through its Square division. Running voice interfaces on local or edge hardware could allow businesses to deploy real-time voice agents at lower hosting costs compared to cloud-hosted audio streams.
At the same time, native voice models remain in early testing stages, leaving cloud-based text and multimodal architectures as the production standard for complex business workflows. Developers looking to compare and integrate state-of-the-art language and reasoning engines can access top AI models cheaply through one unified API at https://apixoai.online, simplifying testing across providers while the ecosystem prepares for the emergence of local speech systems like Dessa.
Source: Block’s new voice model sounds eerily human — TechCrunch AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
